Method and Device for Updating Language Model and Performing Speech Recognition Based on Language Model

1. A method of updating a grammar model used during speech recognition, the method comprising: obtaining a corpus comprising at least one word; obtaining at least one word from the corpus; splitting the obtained at least one word into a prescribed number of segments, wherein the prescribed number is determined based on a size of the grammar model; generating at least one hint for each segment, wherein the at least one hint is for recombining the segments into the at least one word; updating the grammar model by adding each segment to the grammar model with its generated hint; and performing speech recognition based on the updated grammar model.

2. The method of claim 1 , wherein the splitting comprises: determining a word to split based on one or more of an appearance frequency and a length of the obtained at least one word, and splitting the obtained word.

3. The method of claim 1 , wherein the splitting comprises: determining at least one candidate splitting point with respect to the obtained word; obtaining a score of each candidate splitting point based on appearance probability information of a segment based on each candidate splitting point; and determining a splitting point from among the at least one candidate splitting point based on the score, and splitting the obtained word into at least one segment based on the splitting point.

4. The method of claim 1 , wherein the at least one hint represents whether the segment is located at a boundary of the obtained at least one word.

5. The method of claim 1 , wherein the updating of the grammar model comprises: determining appearance probability information of the at least one segment during speech recognition; and updating the grammar model based on the determined appearance probability information.

6. The method of claim 1 , wherein the splitting comprises: splitting the obtained at least one word into segments of phonemes, syllables or morphemes.

7. The method of claim 1 , further comprising: generating a pronunciation variant with respect to the obtained at least one word; obtaining a pronunciation variant corresponding to the generated pronunciation variant from among pronunciation variants included in a pronunciation dictionary; determining appearance probability information of the obtained pronunciation variant during speech recognition; and updating a second grammar model based on the determined appearance probability information.

8. The method of claim 1 , wherein the updating of the grammar model comprises: adding an entry token to the obtained at least one word or the detected segment; and updating a class-based grammar model based on the obtained at least one word or the detected segment comprising the entry token.

9. The method of claim 8 , wherein the updating of the grammar model comprises: determining a class to which the word or the segment belongs; determining whether to add the entry token or an exit token to the word or the segment, and to add the entry token or the exit token to the word or the segment based on a result of the determining; and updating a grammar model of the determined class based on the word or the segment.

10. A method of performing speech recognition, the method comprising: obtaining speech data to perform speech recognition; detecting at least one segment of the speech data by performing speech recognition based on a grammar model comprising appearance probability information of the at least one segment and its hint; and recombining the detected at least one segment into a word based on the hint of the detected at least one segment, wherein the grammar model includes the at least one segment split from the word in a prescribed number, wherein the prescribed number is determined based on a size of the grammar model.

11. The method of claim 10 , further comprising: performing speech recognition so that the word may be detected in a prescribed sequence based on a second grammar model when the word or the detected segment belongs to a first sequence of words.

12. The method of claim 10 , wherein the grammar model comprises a class-based grammar model, and the method comprises: performing speech recognition based on a grammar model of a class to which an entry token belongs when the word or the detected segment comprises the entry token; and performing speech recognition based on the class-based grammar model when the word or the segment comprises an exit token.

13. A non-transitory computer-readable recording medium having recorded thereon a program which, when executed, causes updating a grammar model used during speech recognition, by performing operations comprising: obtaining a corpus comprising at least one word; obtaining at least one word from the corpus; splitting the obtained at least one word into a prescribed number of segments, wherein the prescribed number is determined based on a size of the grammar model; generating a hint for each segment, wherein the at least one hint for recombining the segments into the at least one word; updating the grammar model by adding each segment to the grammar model with its generated hint; and performing speech recognition based on the updated grammar model.